Grid Edge Intelligence (GEI)
- Grid Edge Intelligence is a cyber-physical paradigm that deploys sensing, communication, computation, and AI near grid assets to achieve low-latency, decentralized decision making.
- It utilizes hierarchical architectures—from devices to edge, fog, and cloud—to co-optimize computation, communication, and energy constraints for resilient grid automation.
- Practical implementations include on-device training, local inference for fault detection, renewable asset monitoring, and distributed control that enhance performance and energy efficiency.
Grid Edge Intelligence (GEI) refers to the placement of sensing, communications, computation, control, and AI-driven decision-making near distributed power-system assets rather than only in remote control centers or clouds. In the literature summarized here, GEI appears through on-device training in smart meters, distributed edge-based Fault Location, Isolation and Service Restoration (FLISR), microgrid-side forecasting and droop control on smart meter-concentrators, privacy-preserving coordination of behind-the-meter distributed energy resources (DERs) during black start, edge intelligence for renewable-asset monitoring, and hierarchical edge–fog–cloud computation for smart-grid workloads (Huang et al., 9 Jul 2025, Leniston et al., 2024, Xu et al., 2024, Zheng et al., 18 Aug 2025, Chung et al., 2020, Jackman et al., 13 Oct 2025). The field is therefore not a single method but a systems paradigm in which grid-edge devices become active computational participants in cyber-physical operation.
1. Conceptual scope and distinguishing features
GEI is best understood as a power-system specialization of edge intelligence in which data collection, caching, training, inference, and offloading are performed close to where grid data is produced and where physical actions must be taken (Xu et al., 2020). In smart-grid terms, that shift is motivated by decentralized renewable generation, distributed energy storage, EV charging, smart meters, feeder automation, microgrids, and large IoT sensor populations, all of which create low-latency and high-reliability workloads that are poorly matched to cloud-only architectures (Arcas et al., 2024, Kumar et al., 3 Jun 2026).
A common misconception is that GEI is only on-device inference. The cited literature is broader. It includes on-device model training in a mass-production smart meter, distributed feeder-edge automation for restoration, hierarchical model predictive coordination of customer-owned DERs, offloading and orchestration across edge, fog, and cloud, and non-ML edge automation such as frequency emergency control with coordinated smart outlets (Huang et al., 9 Jul 2025, Leniston et al., 2024, Zheng et al., 18 Aug 2025, Jackman et al., 13 Oct 2025, Xiang et al., 2020). This suggests that GEI spans both learned and rule-based intelligence, provided the intelligence is executed near the operational edge and participates in real control or decision loops.
The surveyed 5G/6G literature sharpens this definition by tying GEI to quantified service requirements. Smart distributed voltage control is associated with roughly $100$ ms latency, smart distributed feeder automation with $2$ ms latency in fault mode, DERs and microgrids with ms latency, real-time fault detection and self-healing with ms latency, and edge-driven data acquisition with ms latency (Kumar et al., 3 Jun 2026). In that sense, GEI is not merely geographically distributed AI; it is latency-bounded, cyber-physical intelligence deployed under operational timing constraints.
2. Architectural patterns and deployment loci
The architectures described in the literature are consistently hierarchical. Device or field layers host sensors, smart meters, relays, inverters, switches, EV chargers, HVAC controllers, and other controllable endpoints; edge layers host feeder, substation, gateway, or concentrator computation; fog layers coordinate groups of edge nodes; cloud layers perform heavier analytics, broader coordination, and model management (Arcas et al., 2024, Xu et al., 2020, Jackman et al., 13 Oct 2025). The 5G/6G survey adds network-side intelligence through MEC, network slicing, O-RAN, and digital twins, making communications part of the GEI stack rather than a passive transport layer (Kumar et al., 3 Jun 2026).
Several papers instantiate this hierarchy in directly grid-relevant forms. The distributed FLISR work for the Irish context places decision logic near feeders and substations, while maintaining central supervisory visibility and policy alignment consistent with ESB Networks and CRU-oriented deployment conditions (Leniston et al., 2024). The microgrid paper places forecasting and droop-control execution on an ARM-based smart meter-concentrator at the point of common coupling, explicitly moving functions out of the far-end control center (Xu et al., 2024). The black-start framework places predictive GEI devices inside residential houses, where they compute flexibility ranges for rooftop PV, battery energy storage, HVAC, and loads, then track utility dispatch without exposing detailed internal asset data (Zheng et al., 18 Aug 2025). The SDEN/CaaS paper generalizes this into a programmable edge–fog–cloud substrate coordinated by SDN, NFV, and 5G URLLC (Jackman et al., 13 Oct 2025).
| GEI instantiation | Edge asset | Primary role |
|---|---|---|
| On-device PV forecasting | Smart meter | Local model training and forecasting |
| Microgrid edge computing | Smart meter-concentrator | Local PV forecasting and droop-control execution |
| Distributed FLISR | Feeder/substation edge nodes | Fault location, isolation, and service restoration |
| GEI-assisted black start | Residential GEI device | Flexibility estimation and dispatch tracking |
| SDEN/CaaS | Edge, fog, and cloud nodes | Task placement under latency and energy constraints |
The architecture literature also broadens GEI beyond fixed utility assets. The MEET framework treats vehicles as mobile edge assets with sensing, communications, computation, storage, and battery-state resources, while the wind-farm study treats UAVs as edge nodes that sense, forecast, optimize routes, and support turbine yaw control (Sun et al., 2022, Chung et al., 2020). These are not conventional distribution-grid controllers, but they show that GEI can include mobile, partially self-powered cyber-physical resources when grid or energy functions are performed close to the asset.
3. Computational mechanisms and control logic
Three computational patterns recur across the literature. The first is local model execution and adaptation. In smart-meter GEI, photovoltaic forecasting is formulated as
with on-device training performed directly on a resource-limited ARM-based meter for XGBoost and LSTM models (Huang et al., 9 Jul 2025). The second is local edge inference and control on low-cost embedded platforms. The microgrid smart meter-concentrator runs LSBoost-based inverter power forecasting and computes and for droop control locally, rather than waiting for remote cloud computation (Xu et al., 2024). The third is distributed or hierarchical optimization across tiers. The SDEN/CaaS formulation places smart-grid tasks across edge, fog, and cloud by minimizing a weighted latency–energy objective,
subject to assignment, capacity, and latency constraints (Jackman et al., 13 Oct 2025).
Training-side mechanisms are similarly diverse. The smart-meter study evaluates full float32 and mixed-precision variants to make on-device LSTM training feasible (Huang et al., 9 Jul 2025). The MEET framework treats federated learning over opportunistic vehicles as a native edge-training mode and explicitly identifies the number of local iterations and the interval between global updates as key scheduling variables (Sun et al., 2022). The broader edge-intelligence survey organizes the space into edge caching, edge training, edge inference, and edge offloading, and emphasizes federated learning, gradient compression, pruning, quantization, knowledge distillation, and device–edge–cloud partitioning as general enablers for resource-constrained edge AI (Xu et al., 2020). The collaborative Edge AI perspective extends this toward federated learning, peer-to-peer collaboration, dynamic clustering, and multi-agent reinforcement learning for decentralized energy systems, though it remains conceptual rather than experimentally validated (Jr et al., 12 May 2025).
GEI also includes non-neural analytics and rule-based evidence fusion. In the Power Distribution Internet of Things, edge smart terminals perform BC-Zscore preprocessing, PCA-based conflict optimization, and Dempster–Shafer evidence fusion over heterogeneous operation, monitoring, and environmental data (Yuan et al., 2022). In emergency frequency control, Grid Sense uses edge smart outlets to detect disturbances, estimate active power loss locally, and make switching decisions using centrally disseminated parameters but without waiting for post-event central commands (Xiang et al., 2020). These examples matter because they show that GEI is not reducible to deep learning; it includes local estimation, evidence fusion, and distributed control logic under severe latency and reliability constraints.
4. Operational domains and representative results
The most mature GEI application domain in the cited literature is resilience-oriented grid automation. The distributed edge FLISR work treats feeder restoration as a distributed control problem executed near field devices, with a communications-aware simulation platform and an explicitly utility-oriented Irish deployment framing (Leniston et al., 2024). The black-start framework extends that resilience orientation to inverter-based restoration, using GEI devices to expose only flexibility ranges and receive dispatch signals, thereby making behind-the-meter DERs usable during bottom-up restoration without direct utility control of private assets (Zheng et al., 18 Aug 2025). Under that framework, restored load-hours increase from $5.36$ MWh at $2$0 GEI penetration to $2$1 MWh at $2$2 GEI penetration (Zheng et al., 18 Aug 2025).
Microgrid and prosumer-edge operation form a second major domain. On-device training in a smart meter shows that XGBoost retraining takes roughly $2$3–$2$4 s with negligible accuracy loss relative to PC training, while LSTM training is feasible but heavier, ranging from $2$5 s to $2$6 s depending on feature length and dataset size; full float32 conversion reduces LSTM wall-clock time by more than half in the reported cases (Huang et al., 9 Jul 2025). The smart meter-concentrator paper shows that embedded forecasting outputs match desktop results to near numerical identity and that inference time is around $2$7 ms per input on the ARM platform, which is sufficient for the paper’s quasi-steady-state microgrid functions (Xu et al., 2024).
Renewable-asset intelligence is another important domain. In wind farms, UAVs perform local wind forecasting using a quantized LSTM and then use that forecast for both inspection-route optimization and turbine yaw-angle decisions (Chung et al., 2020). The reported gains are a $2$8 increase in daily power generation relative to hour-ahead forecasting and about a $2$9 reduction in UAV inspection time relative to the chosen routing baseline (Chung et al., 2020). Although this work is not a distribution-feeder paper, it exemplifies GEI as a local sensing–prediction–optimization–control loop at remote renewable infrastructure.
A fourth domain is large-scale distributed support services. Grid Sense, implemented with smart outlets and reported as deployed in State Grid Jiangsu, detects major disturbances locally and switches responsive loads in about 0 ms after a severe contingency (Xiang et al., 2020). In the reported 1 MW disturbance case, 2 outlets switch off for a total curtailed load of 3 MW, very close to the theoretically required 4 MW, and the frequency nadir is held around 5 Hz (Xiang et al., 2020). This is a particularly clear GEI result because system-level coordination emerges from edge-local decisions parameterized in advance.
5. Communication, energy, privacy, and sustainability constraints
GEI is inseparable from communications engineering. The 5G/6G survey shows that different grid-edge applications impose sharply different requirements on latency, reliability, time synchronization, and data rate, ranging from high-speed differential protection at 6 ms latency with 7 synchronization to smart and intelligent metering at 8 ms latency and edge-driven data acquisition at 9 ms latency (Kumar et al., 3 Jun 2026). The same survey positions MEC, network slicing, AI-driven optimization, digital twins, and O-RAN as core enablers for these differentiated services (Kumar et al., 3 Jun 2026). This suggests that GEI should be treated as a joint computation–communication design problem rather than as local AI deployment in isolation.
Energy and sustainability are likewise explicit design dimensions. MEET argues that fixed 5G/6G infrastructure sized for peak demand is costly and energy intensive, and reports that an infrastructural vehicle carrying Huawei eLTE Rapid has power consumption below 0 W whereas a 4G or 5G base station consumes at least 1 kW (Sun et al., 2022). Its broader claim is that mobility, renewable-powered vehicles, and off-peak charging can spread deployment and operation costs across vastly available vehicles, reduce carbon emissions, and cut electricity bills (Sun et al., 2022). In smart-grid GEI terms, that maps naturally to EV fleets, mobile sensing/compute assets, and joint communication–computation–energy orchestration, even though the original paper is framed around vehicular 6G rather than power systems.
The SDEN/CaaS results make the latency–energy trade-off explicit. For 2 tasks over 3 trials, the reported maximum energy saving relative to cloud-only processing is 4, with a median saving of 5; URLLC integration reduces transmission delay by 6 and jitter by 7; and the reported fault-detection pipeline reaches precision 8, recall 9, F1 0, and 1 ms latency (Jackman et al., 13 Oct 2025). These results reinforce a central GEI point: local or hierarchical execution is valuable only when placement is co-optimized with network capacity and service constraints.
Privacy is a first-order concern rather than a side effect. On-device training in smart meters is motivated by communication cost reduction, agile model updating, and privacy preservation (Huang et al., 9 Jul 2025). The GEI black-start framework exposes only flexibility ranges from household controllers rather than detailed asset information (Zheng et al., 18 Aug 2025). Collaborative Edge AI literature likewise emphasizes federated learning, secure aggregation, homomorphic encryption, secure multi-party computation, and differential privacy as mechanisms for decentralized energy coordination (Jr et al., 12 May 2025). A common misconception is that any edge deployment is automatically privacy-preserving. The literature supports a narrower claim: privacy gains arise when raw data remains local and only constrained summaries, updates, or dispatch interfaces are exchanged.
A forward-looking extension appears in the quantum survey. It argues that quantum sensing is the strongest near-term contributor to grid-edge devices because room-temperature NV-center sensors can improve local observability, while QKD is the strongest near-term communications contribution because it can secure utility links without violating latency requirements (Hoegen et al., 6 Mar 2026). Quantum computing is treated as potentially transformative for optimization and learning, but not yet deployment-ready for direct field-edge use because of size, power, packaging, and interoperability constraints (Hoegen et al., 6 Mar 2026).
6. Limitations, misconceptions, and research directions
The literature is candid about current limitations. Several works are architecture or feasibility studies rather than full end-to-end operational demonstrations. MEET proposes components and case studies but not a unified optimization over mobility, communications, compute, energy, and training/inference placement (Sun et al., 2022). The distributed FLISR paper clearly advances edge restoration, but the available manuscript extract does not expose the full algorithmic and quantitative detail needed for reproducibility (Leniston et al., 2024). The smart-meter study demonstrates feasibility but also shows that deep on-device training remains slow on commodity embedded hardware, especially for LSTM models (Huang et al., 9 Jul 2025). The edge-offloading survey concludes that practical smart-grid pilots remain limited, despite extensive architecture and optimization literature (Arcas et al., 2024).
Another limitation is that adjacent enabling frameworks do not always encode explicit power-system constraints. MEET is highly relevant to GEI because it studies distributed edge AI over heterogeneous, mobile, partially self-powered resources, but it does not model feeder loading, voltages, locational prices, transformer limits, or DER dispatch directly (Sun et al., 2022). The SDEN/CaaS paper is rich in communication–computation co-design, yet its validation is simulation-based and its interoperability with DERMS and legacy SCADA is explicitly left for future work (Jackman et al., 13 Oct 2025). This suggests that GEI research still needs tighter integration between AI placement logic, power-flow and protection models, and real utility operations.
Security, incentives, and governance remain structurally unresolved. MEET explicitly identifies security and privacy guarantees and incentive mechanisms for opportunistic vehicles as future directions (Sun et al., 2022). The collaborative Edge AI perspective emphasizes blockchain, smart contracts, and adaptive governance, but its evidence is largely qualitative (Jr et al., 12 May 2025). The quantum survey adds a different warning: even if QKD strengthens confidentiality, denial-of-service and deployment complexity remain unresolved in large heterogeneous utility networks (Hoegen et al., 6 Mar 2026). Thus GEI should not be equated with autonomous optimization alone; it is also a question of trust boundaries, ownership, and operational accountability.
Several research directions recur across the papers. One is hierarchical, communication-aware orchestration: edge for reactive functions, fog for predictive coordination, cloud for strategic learning and optimization (Jackman et al., 13 Oct 2025). Another is model lifecycle management under changing operating regimes, including concept drift detection and on-demand retraining (Sun et al., 2022). A third is standards-aligned communications support, including MEC, network slicing, O-RAN, digital twins, and 5G/6G service differentiation for grid applications (Kumar et al., 3 Jun 2026). A plausible implication is that GEI will mature not as a single platform, but as a layered operational stack in which cyber-physical intelligence, communications control, privacy preservation, and energy-aware computation are co-designed for specific utility functions rather than deployed as generic edge AI.